a gentle primer on llm explainability
an overview of advances in making large language models more interpretable through dynamic evaluation, statistical methods, and accessible tools.
aitag research
summaries filed under research
microsoft assert tests ai behavior from text descriptions
microsoft's open source assert framework uses ai to turn plain-language rules into scored tests for application-specific ai behavior.
aioptimal transport speeds up wind farm layout design
a new method uses optimal transport to handle symmetry in bayesian optimization, making offshore wind farm layout tuning faster and more reliable.
aidastatformer: hybrid transformer uses stats for das event recognition
a new transformer model uses compact statistical features instead of raw data to classify events in distributed acoustic sensing, cutting data size while keeping accuracy.
aioptimizing vlm reward models with few demonstrations
a test-time method tunes language prompts for vision-language model reward functions using a handful of expert trajectories, reducing false positives without extra training.
airiemannian stochastic optimization for dimension reduction
a new algorithm called smave uses riemannian stochastic gradient ascent on the stiefel manifold for sufficient dimension reduction, avoiding the curse of dimensionality.
aisvm-based quantile regression for extreme inputs
a new svm framework handles quantile regression when covariates are unusually large, using angular components of extreme observations.
aitree search agents optimize molecules for multiple goals
a multi-agent framework uses tree-structured search to coordinate molecular optimization across conflicting objectives, maintaining diverse design paths.
aideliberative curation for multi-agent knowledge bases
a protocol combining reputation-weighted voting and graduated sanctions helps ai agents curate shared knowledge without human-style governance.
aiai models disagree better with byzantine protocol
a new protocol treats ai model disagreement as useful signal, using cognitive personas and validation methods to improve multi-model reasoning.
aihedging on the frontier for few-shot learning
a new approach uses weak monotonicity in benchmark evaluations to improve transfer learning and model selection with few samples.
aiqasm-eval dataset for llms on openqasm-3
a new dataset trains and evaluates large language models on openqasm-3 programs with advanced hardware-oriented features beyond simple quantum circuits.
aiagent logic guides enterprise ai beyond llms
ibm research shows that adding software primitives like knowledge graphs and program analysis to ai agents improves performance and cuts costs in enterprise workflows.
aideep learning predicts hip forces from gait kinematics
a new benchmark uses deep learning to estimate hip muscle forces and joint moments directly from walking data, tested on healthy adults and patients.
aibatched bandits with 1-bit feedback
a new study examines stochastic linear bandits where the learner gets only one bit of feedback per batch of actions, revealing fundamental limits and near-optimal algorithms.
aillms without deep neural networks: closed-form global optimum
a new architecture replaces deep neural networks in llms by finding the global optimum in one step, removing the need for iterative training.
aiphysically grounded diagram generation from text
a neuro-symbolic pipeline generates physics diagrams from text by enforcing physical laws through a scene graph, solver, and verification loop.
aiuncertainty-aware expert advice for safer self-driving cars
a new reinforcement learning method uses expert guidance only when the agent is uncertain, reducing crashes in simulated autonomous driving.
aiunicorn scales time series forecasting across datasets
a new framework uses a latent prototype codebook to model channel correlations without being tied to specific channel identities, enabling multi-dataset pretraining and strong few-shot transfer.
aiwhy world models need physical viability
a new paper argues that world models for embodied ai must represent physical structure to answer intervention queries, not just predict observations.
ai